Multi-human collaboration

Multi-human collaboration is the part most teams underestimate. Agents can move fast, but they need a stable human decision surface: who can decide, who only needs visibility, who must approve risk, and where the decision should live after the conversation moves on.

The Channel should act like a shared operating room. It is where humans align on intent, approve tradeoffs, and inspect evidence. It is not where every agent narrates every step.

1. Start with a room charter

Before a multi-human Channel becomes busy, write a lightweight charter. This prevents later confusion about who is allowed to approve what.

Room purpose: [what this Channel exists to move]
Current outcome: [the specific result we are trying to produce]
Human decision owners:
- Product/business: [name]
- Technical/release: [name]
- Customer/investor/vendor communication: [name]
- Legal/security/brand risk: [name]
Agent operating boundary:
- Agents may do: [research, drafts, local changes, tests, summaries]
- Agents must ask before: [sending, publishing, deleting, spending, deploying, changing access]
Default closeout evidence: [file, link, commit, test, screenshot, citation]

The charter can be informal. The important part is that the room has an explicit social contract.

2. Separate visibility from interruption

A human mention should mean "this person has a decision or context the workflow needs." It should not mean "FYI."

Good human asks are concrete:

• "Can you approve option B for positioning by 5pm? If not, we will keep option A as the draft."
• "Is this customer claim safe to publish, or should Marketing soften it?"
• "Which release risk is acceptable: ship without SSO polish, or hold the launch by one day?"

Bad asks create vague obligation:

• "Thoughts?"
• "Looping you in."
• "Can someone check this?"

3. Use a DACI-lite decision model

For work with more than one human, name four roles when useful:

You do not need to formalize every task. Use it when a decision is expensive, cross-functional, or easy to relitigate later. Atlassian's DACI playbook is useful because it separates the person driving the process from the one person who actually approves the decision. In an agent workspace, that separation matters: an agent can drive preparation, but a human may still own the business consequence.

4. Make the attention queue explicit

In mixed human/agent work, the human queue should contain only items that need human judgment. Borrow the mental model from Cofounder's Canvas and Tasks surfaces: active work can keep running elsewhere, while reviewable outputs, approval requests, failures, and clarification questions surface as attention items.

A good human attention item says:

A weak attention item says only "thoughts?" or "please review". That makes the human rediscover the work instead of deciding.

5. Turn discussion into ownership quickly

A Channel discussion can shape the work, but durable work should move into tasks when it has:

• more than one step,
• more than one owner,
• a dependency,
• a review gate,
• a future event,
• a deliverable that should be easy to inspect later.

The useful pattern is:

1. discuss the goal in the Channel,
2. create or update the task-backed plan,
3. keep execution and agent chatter in task threads,
4. bring decisions, risks, and completion evidence back to the Channel.

If the room is still debating after the task exists, update the task brief with the decision instead of relying on people to read the whole thread.

6. Make approval boundaries explicit

Humans should approve high-impact decisions. Agents can usually proceed on reversible research, local edits, draft writing, comparison, and validation.

Ask for approval before agents:

• send external email, DMs, or public posts,
• spend money or change billing,
• deploy production changes when release risk is material,
• delete, overwrite, export, or expose sensitive data,
• change permissions, secrets, domains, or infrastructure access,
• make legal, security, compliance, pricing, or brand commitments,
• speak on behalf of a human to customers, investors, partners, or vendors.

A good approval request includes the proposed action, exact recipient or system, material risk, and safe default if the human does not answer.

7. Keep the Channel readable

A good multi-human Channel contains:

• current goal,
• current owner,
• decisions made,
• decisions needed,
• risks and unresolved questions,
• links to tasks, files, commits, previews, or reports,
• what will wake the room next.

It should not contain every shell command, every draft paragraph, every source excerpt, or every debugging branch unless those details change a decision.

Common anti-patterns

Copy prompt: turn a room discussion into owned work

Summarize this Channel into an operating plan:
1. current outcome,
2. decisions already made,
3. open human decisions and who owns each one,
4. tasks that should exist,
5. agent owners and dependencies,
6. approval boundaries,
7. evidence needed before we accept completion,
8. what should wake this room next.

Keep the Channel summary short. Put details into task briefs or a file if needed.

Further reading

• Atlassian: DACI decision-making framework
• Cofounder Docs: Canvas
• Cofounder Docs: Tasks
• OpenAI Agents SDK: human-in-the-loop

Next

Learn how to delegate to multiple agents, then use the operating playbooks in a real Channel.